Investigating Algorithmic Bias in YouTube Shorts

Diwash Poudel; Mert Can Cakmak; Nitin Agarwal

REVIEW 5 major objections 5 minor 1 cited by

Investigating Algorithmic Bias in YouTube Shorts

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims YouTube Shorts' recommendation algorithm systematically replaces politically sensitive content with entertainment within the first recommendation step, regardless of watch time.

desk verdict Solid data effort, but missing control seeds and inferential statistics make the 'political bias' headline an over-reach. read the letter →

arxiv 2507.04605 v1 pith:D2XSLZED submitted 2025-07-07 cs.SI

Mert Can Cakmak , Nitin Agarwal , Diwash Poudel This is my paper

classification cs.SI

keywords YouTubeShortsalgorithmicbiasrecommendationdriftcontentemotionanalysiswatch-timepopularitypolitical

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that YouTube Shorts' recommendation algorithm shows a systematic bias: within the first recommendation step it drops politically sensitive topics and shifts toward entertainment, joyful or neutral emotional tone, and high-engagement videos. The authors audited 685,842 videos collected from seed videos about the South China Sea dispute, the 2024 Taiwan presidential election, and general Shorts, classifying each by relevance, topic, and emotion with a validated generative-AI model. They find relevance to the original political topic falls to near zero at depth 1 and that varying simulated watch time (3, 15, or 60 seconds) does not restore topical alignment. If correct, the finding matters because it describes how a platform serving over two billion users shapes exposure to political information.

What carries the argument

The core object is a recommendation chain: each seed Short is opened in a fresh, logged-out browser session and the next 50 recommended Shorts are recorded, with 3-, 15-, or 60-second dwell times simulating watch-time. Relevance, topic category (politics, non-entertainment, entertainment), and emotion (joy, sadness, anger, neutral, fear) are assigned by GPT-4o from each video's title and transcript, after validation against benchmark datasets for emotion, topic, and relevance. This setup turns an opaque proprietary recommender into measured drift curves of relevance, topic, emotion, and engagement as functions of recommendation depth.

What would settle it

Run a controlled, logged-in session seeded with a South China Sea or Taiwan-election Short, record the first ten recommendations while the user actually watches for 15 or 60 seconds, and compute semantic relevance to the seed topic; if mean relevance remains substantially above near-zero at depth 5 in this setting, the claimed immediate, watch-time-independent drift is contradicted for real user conditions.

Watch

Extended reading notes

Core claim

The central claim is that YouTube Shorts' recommendation chain diverges from politically sensitive seed content almost immediately and does so consistently across watch-time conditions. For both the South China Sea and Taiwan election datasets, seed videos are highly relevant and predominantly political, but by the first recommended video relevance scores approach zero and entertainment content becomes the majority, with joy/happiness rising and anger or fear declining. Engagement metrics jump after the first recommendation, indicating that highly viewed and liked videos are disproportionately promoted, and in the 60-second condition periodic sponsored content appears beyond depth 10. The authors conclude that watch time does not reduce algorithmic bias, and that the platform systematically prefers emotionally positive, engagement-optimized content over politically sensitive material.

Load-bearing premise

The watch-time results assume that pausing for 3, 15, or 60 seconds in an automated browser is perceived by YouTube as a genuine watch-time signal, but the study offers no evidence that the platform registers these pauses as engagement.

Editorial extensions

If this is right

A user who starts with a politically sensitive Short will, within one recommendation step, mostly see entertainment content rather than related political coverage.
Longer viewing does not counteract the drift, so the common assumption that engagement signals keep a feed on-topic is not supported for Shorts.
High-engagement videos are systematically promoted, which can reinforce popularity bias and crowd out niche or serious topics.
In longer viewing sessions, sponsored or promotional content is inserted periodically after about the tenth recommendation, suggesting monetized content is timed to sustained attention.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the near-zero relevance at depth 1 may reflect Shorts' feed design, which could be dominated by global or trending content rather than topic affinity; one test would be whether entertainment seeds also drift topically.
Beyond the paper, the watch-time conclusion depends on the untested assumption that an automated browser pause is registered by YouTube as genuine watch-time, so a logged-in real-user study is needed before generalizing the 'watch time does not reduce bias' claim.
Beyond the paper, emotion labels come from titles and transcripts, so visually or musically affective Shorts with little text may be misclassified; a multimodal annotation could change the measured emotional drift.
Beyond the paper, because sessions are logged-out and started from clean browsers, the results describe cold-start recommendations, not the personalized feeds most users experience.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a real data-collection effort with a transparent pipeline, but the central claim that YouTube Shorts' algorithm is specifically biased against political content doesn't survive the missing control condition. The paper's own general-content baseline is entertainment-dominated at depth 0, so the observed drift from political seeds to entertainment could just be generic popularity bias affecting any niche cold-start topic.

What's genuinely new: the scale (685k videos), the three watch-time conditions, and the South China Sea/Taiwan seed set. The methodology is honestly described—GPT-4o is benchmark-validated (though with moderate accuracy), and the authors are clear that they measure observed recommendations, not internal parameters. That is the right way to audit a proprietary system.

Where it gets soft. First, every drift claim rests on mean curves with shaded standard deviations in Figures 1–4. There are no confidence intervals, significance tests, or effect sizes, so we can't tell which of the many apparent trends are meaningful. Second, the watch-time simulation is unverified: pausing in Selenium for 3, 15, or 60 seconds in a fresh logged-out browser is not shown to be perceived by YouTube as watch-time engagement. If it isn't, RQ3 collapses. Third, the novelty statement is overstated—the authors' own [10] and [12] already report drift to entertainment and topic dynamics in YouTube Shorts. The new specific combination is incremental. Fourth, the classification accuracy (emotion 66.95%, relevance 78.73%) is moderate and validated only on general benchmarks, not on Shorts text.

The most serious issue is the one the stress-test note flags: there is no matched control seed for a non-political, non-entertainment topic. Science, education, or hobby seeds with similar length and engagement would tell you whether the algorithm avoids politics specifically or just regresses to the platform's entertainment-heavy base rate. The paper even concedes in Section 4.2 that the drift 'likely reflects engagement optimization and popularity bias'—a generic mechanism. Without the control, 'systematic avoidance of political topics' is an over-read.

Bottom line: the dataset is worth having and the paper is written clearly, but the headline interpretation is not established. A serious referee should see it—the design can be fixed with control seeds, proper statistics, and a softened claim. I'd bring it to a reading group as a case study in how easy it is to over-interpret observational recommendation traces.

Referee Report

5 major / 5 minor

Summary. The paper reports an algorithm audit of YouTube Shorts recommendations. The authors seed fresh, logged-out browsing sessions with videos on the 2024 Taiwan presidential election, the South China Sea dispute, and a general content category; collect recommendation chains up to depth 50 under three simulated watch-time pauses (3s, 15s, 60s); classify 685,842 videos using GPT-4o on relevance, topic, and emotion; and report drift from politically sensitive seeds toward entertainment and positive affect, a preference for high-engagement videos, and no reduction in bias with longer watch time. The paper claims to be the first comprehensive analysis of algorithmic drift in YouTube Shorts along these dimensions.

Significance. If the central claims were established, this would be a useful contribution to the algorithm-auditing literature: the study is large in scale, measures observed recommendations rather than hypothetical model outputs, compares three watch-time conditions, and uses multi-dimensional labeling with an explicit benchmark-based model selection. The controlled fresh-session collection design is a strength, as is the authors' transparency about the validation benchmarks. However, the current evidence does not support the headline political-bias conclusion: the design lacks a matched non-political control, the statistical analysis is almost entirely descriptive, the classification is validated only on domain-mismatched benchmarks, and the watch-time manipulation is not validated as a genuine engagement signal. The significance is therefore potential rather than realized.

major comments (5)

[Section 3.1, Table 1; Section 4.2; Abstract] The central claim that YouTube Shorts 'systematically' deprioritizes politically sensitive content cannot be separated from generic popularity bias because the only comparison condition is not a matched control. The General Content set is an aggregate of 15 categories and is already entertainment-dominated at depth 0 (Figures 2g-2i), so the observed drift from political seeds to entertainment is equally consistent with cold-start regression toward the dominant Shorts distribution in fresh, logged-out sessions. A matched control of non-political, non-entertainment seeds (e.g., science or education) with similar length, recency, and initial engagement is required to attribute the drift to topic sensitivity. Section 4.2's own concession that the effect 'likely reflects engagement optimization and popularity bias' admits that the mechanism has not been identified as specifically political. This is a load-bearing gap for the abstract's conclusion about politically sensitive content.
[Sections 4.1-4.4, Figures 1-4] All drift claims rest on visual inspection of mean depth curves. The paper reports no confidence intervals, significance tests, effect sizes, or multiple-comparison corrections; standard deviations are shown as additional curves rather than as uncertainty bands around the means. In particular, the claim that 'watch time does not meaningfully impact topic relevance' (Section 4.1) and the detailed statements about differences among the 3s, 15s, and 60s conditions (Sections 4.2 and 4.4) are asserted from plotted means without any statistical comparison. The strength of the headline conclusions--'consistent drift,' 'systematic preference,' 'does not reduce algorithmic bias'--is not matched by the reported evidence.
[Section 3.2, Table 4; Section 4.1] The classifier used for all downstream analyses is validated only on generic benchmark datasets that do not match the YouTube Shorts distribution. Emotion accuracy is 66.95% and relevance accuracy is 78.73%, and relevance is precisely the measure supporting the 'approach[ing] near-zero values at depth 1' claim. Because relevance is judged against a broadly formulated topic ('the South China Sea dispute' or 'the Taiwan election') rather than against the specific seed video, and because no domain-specific validation or error analysis on Shorts titles/transcripts is provided, classifier noise could plausibly inflate or even generate the apparent sharp relevance drop. This is a measurement-validity concern for the paper's primary drift metric.
[Section 3.1, Section 5 (RQ3)] The watch-time manipulation is not validated as a genuine engagement signal. Pausing a Selenium browser for 3, 15, or 60 seconds is assumed to be interpreted by YouTube as watch time, but no evidence is offered that the platform registers these pauses as dwell time, and no post-hoc check verifies that the recommendation outputs differ in a way attributable to the pause rather than to elapsed time or other confounds. Section 3.1 only says the delays were incorporated 'to better simulate natural user behavior,' which does not establish signal validity. Since RQ3 and the conclusion that 'watch time does not reduce algorithmic bias' depend entirely on this assumption, the watch-time dimension of the study is unsupported.
[Section 4.4, Figure 4] The engagement analysis conflates the properties of recommended videos with the algorithm's promotion decisions. Showing that mean logged views, likes, and comments rise from depth 0 to depth 1 does not demonstrate that the algorithm 'disproportionately promotes' high-engagement content without comparing against a suitable base rate of Shorts engagement or otherwise adjusting for confounds. Seed videos were collected through keyword searches and may be systematically less viral than the general Shorts population, making the observed depth-0 to depth-1 rise partly a selection artifact. The claim of popularity bias therefore needs a baseline comparison or an appropriate control, not just a within-chain trend.

minor comments (5)

[Table 2 and Section 3.1] Table 2's Total column sums to 681,300, while the text reports 685,842 videos 'including both root and recommended videos'; the difference is exactly the number of root videos, so the table appears to omit root videos from the Total entries. The column header and the text should be reconciled.
[Figure captions, Figures 1-4] The captions describe standard deviations as plotted curves, which is visually confusing; showing shaded confidence bands or reporting per-depth summary statistics in tables would make the uncertainty clearer and would also strengthen the paper's inferential claims.
[Sections 4.2 and 4.4] The paper repeatedly attributes observed patterns to 'ads or sponsored content' (e.g., Figures 2c, 3c, 4c), but no detection method for ads or sponsored videos is described. The authors should either operationalize this category or avoid making specific claims about promotional insertions.
[Section 3.1] The general content condition includes 'News & Politics' as one of its 15 categories, so it is not a clean non-political baseline; the near-zero political share at depth 0 in Figure 2 should be interpreted and reported with this composition in mind.
[References] A few references are missing publication venues or complete author information (e.g., [11], [17], [25]); the bibliography should be brought into a consistent format.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the drift measurements are externally labeled observations, not fitted or definitional outputs.

full rationale

The paper's central claims (drift from political to entertainment content, emotion shift toward joy/neutral, engagement amplification, and watch-time insensitivity) are direct empirical measurements of recommendation outputs collected from fresh, logged-out Selenium sessions. The labels are produced by GPT-4o, which was selected by validation on external benchmark datasets (GoEmotions, DailyDialog, BBC News, News Category, MS MARCO, WikiQA) reported in Tables 3-4; no classifier weight or threshold is fitted to the YouTube Shorts data whose drift is then reported, so the 'prediction' is not forced by construction. The relevance, topic, and emotion categories are operational definitions, and the finding that depth-0 political seeds lose relevance at depth 1 is an observed property of the recommendation chain, not an analytic consequence of the definition. The paper's own Section 4.2 notes that the drift 'likely reflects engagement optimization and popularity bias,' which is a substantive interpretation rather than a circular reduction to the seed keywords. Self-citations (refs [8]-[10], [12], [20]-[21]) appear in the literature review and methodology for prior findings and transcript tooling, but the load-bearing conclusion does not depend on accepting those prior results: the current study's data collection and external model validation stand independently. The absence of a matched non-political, non-entertainment control is a validity and generalizability concern about confounding with base-rate popularity bias, not a circularity in the derivation chain. No equation or fitted parameter is reused as an output, so score 0 is appropriate.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims depend on the reliability of GPT-4o labels for YouTube Shorts content, on the fidelity of simulated watch-time as a signal to YouTube, and on the assumption that logged-out browser instances are neutral. None of these premises is independently verified in the paper.

free parameters (2)

watch-time thresholds = 3s, 15s, 60s
Three viewing durations chosen by hand to simulate engagement levels; central to RQ3, but no evidence that YouTube perceives these pauses as watch-time.
recommendation depth limit = 50
Maximum scroll depth chosen to balance comprehensiveness and runtime; all analyses use depth-50 chains, and the exact total count implies every seed produced exactly 50 recommendations.

assumptions (4)

domain assumption GPT-4o classifications are valid for YouTube Shorts content
Model validated on six generic benchmarks (GoEmotions, DailyDialog, BBC News, News Category, MS MARCO, WikiQA), not on YouTube Shorts data; emotion accuracy 66.95%, topic 85.15%, relevance 78.73% (Table 4).
domain assumption Pausing 3s/15s/60s in a headless browser is interpreted by YouTube as genuine watch-time signals
Section 3.1 states manual scrolling delays were incorporated to simulate natural behavior, but no evidence that the platform registers these pauses as engagement.
domain assumption Recommendation chain depth ordering reflects YouTube's recommendation priority
Assumes depth 1 through 50 corresponds to algorithmic ordering; YouTube may interleave ads, mixed content, and personalized signals not captured by a logged-out instance.
domain assumption A fresh WebDriver with no cookies and no logged-in user is a neutral environment
Section 3.1 claims neutrality, but logged-out YouTube still personalizes by IP, location, and device, so neutrality is not established.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Investigating Algorithmic Bias in YouTube Shorts." pith.science (2026). https://pith.science/paper/D2XSLZED

@misc{pith2026250704605,
  author       = {Pith},
  title        = {Pith review of: Investigating Algorithmic Bias in YouTube Shorts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2XSLZED}},
  note         = {Machine review of arXiv:2507.04605}
}

read the original abstract

The rapid growth of YouTube Shorts, now serving over 2 billion monthly users, reflects a global shift toward short-form video as a dominant mode of online content consumption. This study investigates algorithmic bias in YouTube Shorts' recommendation system by analyzing how watch-time duration, topic sensitivity, and engagement metrics influence content visibility and drift. We focus on three content domains: the South China Sea dispute, the 2024 Taiwan presidential election, and general YouTube Shorts content. Using generative AI models, we classified 685,842 videos across relevance, topic category, and emotional tone. Our results reveal a consistent drift away from politically sensitive content toward entertainment-focused videos. Emotion analysis shows a systematic preference for joyful or neutral content, while engagement patterns indicate that highly viewed and liked videos are disproportionately promoted, reinforcing popularity bias. This work provides the first comprehensive analysis of algorithmic drift in YouTube Shorts based on textual content, emotional tone, topic categorization, and varying watch-time conditions. These findings offer new insights into how algorithmic design shapes content exposure, with implications for platform transparency and information diversity.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Algorithmic Audit of Personalisation Drift in Polarising Topics on TikTok
cs.IR 2026-03 conditional novelty 6.0 of 10

TikTok personalises strongly, neutralises climate/vaccine/flat-earth content toward safe neutral topics, but sustains and often stance-reinforces US politics—with a tilt toward the oppose stance when both sides are seeded.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages · cited by 1 Pith paper

[10]

Social Network Analysis and Mining 14(1), 1–42 (2024)

Cakmak, M.C., Agarwal, N., Oni, R.: The bias beneath: analyzing drift in youtube’s algorithmic recommendations. Social Network Analysis and Mining 14(1), 1–42 (2024)

work page 2024
[12]

In: International Confer- ence on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, pp

Cakmak, M.C., Agarwal, N., Dagtas, S., Poudel, D.: Unveiling bias in youtube shorts: Analyzing thumbnail recommendations and topic dynamics. In: International Confer- ence on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, pp. 205–215 (2024). Springer

work page 2024
[1]

Shorts on the Rise: Assessing the Effects of YouTube Shorts on Long-Form Video Content

Rajendran, P.T., Creusy, K., Garnes, V.: Shorts on the rise: Assessing the effects of youtube shorts on long-form video content. arXiv preprint arXiv:2402.18208 (2024)

work page Pith review arXiv 2024
[2]

Aubin, C., Liedke, J.: Social Media and News Fact Sheet

St. Aubin, C., Liedke, J.: Social Media and News Fact Sheet. Pew Research Center. Accessed: October 3, 2024 (2024). https://www.pewresearch.org/journalism/fact-sheet/ social-media-and-news-fact-sheet/

work page 2024
[3]

Shorts vs. Regular Videos on YouTube: A Comparative Analysis of User Engagement and Content Creation Trends

Violot, C., Elmas, T., Bilogrevic, I., Humbert, M.: Shorts vs. regular videos on youtube: A comparative analysis of user engagement and content creation trends. arXiv preprint arXiv:2403.00454 (2024)

work page Pith review arXiv 2024
[4]

Michigan Journal of International Law 45(1), 93–154 (2024)

Seo, Y.: Power shift, the south china sea dispute, and the role of international law. Michigan Journal of International Law 45(1), 93–154 (2024)

work page 2024
[5]

The New York Times

The New York Times: How Taiwan’s Election Fits Into the Island’s Past, and Its Future. The New York Times. Accessed: 2024-06-30 (2024). https://www.nytimes.com/2024/01/ 14/world/asia/taiwan-election-china-lai-ching-te.html

work page 2024
[6]

In: Boratto, L., Faralli, S., Marras, M., Stilo, G

Kirdemir, B., Kready, J., Mead, E., Hussain, M.N., Agarwal, N.: Examining video rec- ommendation bias on youtube. In: Boratto, L., Faralli, S., Marras, M., Stilo, G. (eds.) Advances in Bias and Fairness in Information Retrieval, pp. 106–116. Springer, Cham (2021)

work page 2021

Show all 28 references

[7]

Applied Network Science 10(1), 1–38 (2025)

Cakmak, M.C., Agarwal, N.: Influence of symbolic content on recommendation bias: ana- lyzing youtube’s algorithm during taiwan’s 2024 election. Applied Network Science 10(1), 1–38 (2025)

2025
[8]

In: Proceedings of the International Conference on Advances in Social Networks Analysis and Mining, pp

Cakmak, M.C., Okeke, O., Onyepunuka, U., Spann, B., Agarwal, N.: Analyzing bias in rec- ommender systems: A comprehensive evaluation of youtube’s recommendation algorithm. In: Proceedings of the International Conference on Advances in Social Networks Analysis and Mining, pp. 7...

2024
[9]

In: Cherifi, H., Rocha, L.M., Cherifi, C., Donduran, M

Cakmak, M.C., Okeke, O., Onyepunuka, U., Spann, B., Agarwal, N.: Investigating bias in youtube recommendations: Emotion, morality, and network dynamics in china-uyghur content. In: Cherifi, H., Rocha, L.M., Cherifi, C., Donduran, M. (eds.) Complex Networks & Their Applications...

2024
[11]

Yang, C.: Bias in short-video recommender systems: user-centric evaluation on tiktok (2022) 14

2022
[13]

Hunnego, M.: Exploring the consistency and variability of algorithmic filter bubbles: A comparative analysis of instagram reels and tiktok. B.S. thesis, University of Twente (2024)

2024
[14]

In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp

Zhang, Y., Bai, Y., Chang, J., Zang, X., Lu, S., Lu, J., Feng, F., Niu, Y., Song, Y.: Lever- aging watch-time feedback for short-video recommendations: A causal labeling framework. In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management...

2023
[15]

PBS News

Klepper, D., Wu, H.: How Taiwan preserved election integrity by fighting back against disinformation. PBS News. Accessed: 2024-06-30 (2024). https://tinyurl.com/46n27r5w

2024
[16]

https://www.atlanticcouncil.org/ programs/digital-forensic-research-lab/

Atlantic Council: Digital Forensic Research Lab. https://www.atlanticcouncil.org/ programs/digital-forensic-research-lab/. Accessed: 2025-02-25 (2025)

2025
[17]

https://cil.nus.edu.sg/research/ocean-law-policy/south-china-sea/

Centre for International Law, National University of Singapore: Ocean Law and Pol- icy – South China Sea. https://cil.nus.edu.sg/research/ocean-law-policy/south-china-sea/. Accessed: 2025-02-25 (2025)

2025
[18]

https://entreresource.com/ youtube-video-categories-full-list-explained-and-which-you-should-use/

EntreResource: YouTube Video Categories – Full List Explained and Which You Should Use in 2023. https://entreresource.com/ youtube-video-categories-full-list-explained-and-which-you-should-use/. Accessed: 2025-02-25 (2023)

2023
[19]

Streamers: Youtube Scraper. APIFY. Accessed: 2024-01-10. https://apify.com/streamers/ youtube-scraper

2024
[20]

In: 2023 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp

Cakmak, M.C., Okeke, O., Spann, B., Agarwal, N.: Adopting parallel processing for rapid generation of transcripts in multimedia-rich online information environment. In: 2023 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp. 832–837 (2023)...

2023
[21]

In: 2024 IEEE Interna- tional Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp

Cakmak, M.C., Agarwal, N.: High-speed transcript collection on multimedia platforms: Advancing social media research through parallel processing. In: 2024 IEEE Interna- tional Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp. 857–860 (2024). https://doi.org...

2024
[22]

IEEE Engineering Management Review 52(2), 153– 164 (2024) https://doi.org/10.1109/EMR.2024.3353338

Joosten, J., Bilgram, V., Hahn, A., Totzek, D.: Comparing the ideation quality of humans with generative artificial intelligence. IEEE Engineering Management Review 52(2), 153– 164 (2024) https://doi.org/10.1109/EMR.2024.3353338

2024
[23]

https://arxiv.org/abs/2005.00547

Demszky, D., Movshovitz-Attias, D., Ko, J., Cowen, A., Nemade, G., Ravi, S.: GoEmotions: A Dataset of Fine-Grained Emotions (2020). https://arxiv.org/abs/2005.00547

2020 arXiv
[24]

https://arxiv.org/abs/1710.03957

Li, Y., Su, H., Shen, X., Li, W., Cao, Z., Niu, S.: DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset (2017). https://arxiv.org/abs/1710.03957

2017 arXiv
[25]

https://www.kaggle.com/datasets/bhavikjikadara/ bbc-news-articles

Jikadara, B.: BBC News Articles. https://www.kaggle.com/datasets/bhavikjikadara/ bbc-news-articles. Kaggle (2023)

2023
[26]

https://arxiv.org/abs/2209.11429

Misra, R.: News Category Dataset (2022). https://arxiv.org/abs/2209.11429

2022 arXiv
[27]

Nguyen, T., Rosenberg, M., Song, X., Gao, J., Tiwary, S., Majumder, R., Deng, L.: Ms marco: A human-generated machine reading comprehension dataset (2016)

2016
[28]

In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp

Yang, Y., Yih, W.-t., Meek, C.: Wikiqa: A challenge dataset for open-domain question answering. In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 2013–2018 (2015) 15

2015

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [10]

Social Network Analysis and Mining 14(1), 1–42 (2024)

Cakmak, M.C., Agarwal, N., Oni, R.: The bias beneath: analyzing drift in youtube’s algorithmic recommendations. Social Network Analysis and Mining 14(1), 1–42 (2024)

work page 2024

[2] [12]

In: International Confer- ence on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, pp

Cakmak, M.C., Agarwal, N., Dagtas, S., Poudel, D.: Unveiling bias in youtube shorts: Analyzing thumbnail recommendations and topic dynamics. In: International Confer- ence on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, pp. 205–215 (2024). Springer

work page 2024

[3] [1]

Shorts on the Rise: Assessing the Effects of YouTube Shorts on Long-Form Video Content

Rajendran, P.T., Creusy, K., Garnes, V.: Shorts on the rise: Assessing the effects of youtube shorts on long-form video content. arXiv preprint arXiv:2402.18208 (2024)

work page Pith review arXiv 2024

[4] [2]

Aubin, C., Liedke, J.: Social Media and News Fact Sheet

St. Aubin, C., Liedke, J.: Social Media and News Fact Sheet. Pew Research Center. Accessed: October 3, 2024 (2024). https://www.pewresearch.org/journalism/fact-sheet/ social-media-and-news-fact-sheet/

work page 2024

[5] [3]

Shorts vs. Regular Videos on YouTube: A Comparative Analysis of User Engagement and Content Creation Trends

Violot, C., Elmas, T., Bilogrevic, I., Humbert, M.: Shorts vs. regular videos on youtube: A comparative analysis of user engagement and content creation trends. arXiv preprint arXiv:2403.00454 (2024)

work page Pith review arXiv 2024

[6] [4]

Michigan Journal of International Law 45(1), 93–154 (2024)

Seo, Y.: Power shift, the south china sea dispute, and the role of international law. Michigan Journal of International Law 45(1), 93–154 (2024)

work page 2024

[7] [5]

The New York Times

The New York Times: How Taiwan’s Election Fits Into the Island’s Past, and Its Future. The New York Times. Accessed: 2024-06-30 (2024). https://www.nytimes.com/2024/01/ 14/world/asia/taiwan-election-china-lai-ching-te.html

work page 2024

[8] [6]

In: Boratto, L., Faralli, S., Marras, M., Stilo, G

Kirdemir, B., Kready, J., Mead, E., Hussain, M.N., Agarwal, N.: Examining video rec- ommendation bias on youtube. In: Boratto, L., Faralli, S., Marras, M., Stilo, G. (eds.) Advances in Bias and Fairness in Information Retrieval, pp. 106–116. Springer, Cham (2021)

work page 2021